Research · April 1980
Kunihiko Fukushima described a multilayer network for pattern recognition that tolerates shifts in position.
Its local features and pooling layers precede the convolutional networks that later became the basis of computer vision.
The network was trained by self-organisation rather than by backpropagation.
Biological Cybernetics · Published April 1980
Both recognise images through layers of local features; the later one ran on much larger data and on GPUs.
The 1980 neocognitron is the same author's direct continuation of this work.
Cognitron: A self-organizing multilayered neural networkConvolution and weight sharing come from the neocognitron, but here they are trained by gradient.
Backpropagation Applied to Handwritten Zip Code RecognitionFukushima writes that the neocognitron, once self-organised, has a structure like Hubel and Wiesel's hierarchy model: its S-cells resemble simple cells and its C-cells complex ones.
Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position